AI Detection
What does an AI detector score actually mean?
A percentage, a highlighted passage and a verdict answer different questions. Read them without confusing a signal with proof.
An AI detector gives you evidence to review, not a record of who typed each sentence. Before interpreting a number such as 40%, ask: 40% of what? The answer depends on the product and the field in its report.
Three numbers that should not be confused
A document-level probability expresses a model's confidence in a classification. An estimated AI-text share describes how much of the analysed text the system considers AI-like. Highlight coverage measures the text actually coloured in the interface. These are not interchangeable measures of accuracy.
A hypothetical report might highlight 30 words in a 100-word passage. That is 30% highlight coverage. It does not establish that exactly 30 words were generated by AI, nor make authorship 30% certain. This example explains a denominator; it is not a Certaffy benchmark.
Different detectors define their scores differently
Turnitin's reporting guide describes a percentage of qualifying prose identified as likely AI-generated. GPTZero's interpretation guide describes document-classification probabilities. A score of 40 in one interface is not automatically equivalent to 40 in the other.
Compare definitions, supported inputs and evaluation methods before comparing numbers. Certaffy has not established superiority over those products in a shared, independent benchmark. This explanation is not a league table.
False positives and missed AI are different errors
A human-document false positive occurs when a human-written document is classified as AI. A false highlight occurs when human-written text is coloured as suspicious. A system can avoid false highlights by missing much of the AI text, so ask for recall as well as precision.
Check the test's sample size and types of writing. A small test of unedited AI essays cannot establish performance on business emails, edited drafts or scanned PDFs. The RAID research benchmark found weaknesses under changes in domain, generation settings and adversarial edits. Its results concern that study, not a current ranking of every detector.
Read a report in this order
- Confirm that the extracted text matches your document.
- Read short-text, unsupported-input and confidence warnings first.
- Identify whether the percentage is a probability, estimated share or coverage.
- Read highlighted passages in context, including citations and surrounding sentences.
- Consider draft history, notes and the author's explanation when appropriate.
A withheld score is not proof of human authorship. An unhighlighted sentence is not a guarantee either. If a result will affect a student or employee, use the relevant review process rather than the score alone.
Explore Certaffy's free AI writing check and our fair academic review checklist. Features and limits vary by plan; a demonstration is not an accuracy guarantee.
Prepared with AI assistance for Certaffy. Source links checked on September 26, 2026. Product definitions may change.